BactInt: A domain driven transfer learning approach for extracting inter-bacterial associations from biomedical text.

Area of Science:

• Life Sciences
• Bioinformatics
• Medical Informatics

Background:

• Bacterial associations influence ecosystem health, including the human body.
• Biomedical text is a key source for identifying inter-bacterial relationships.
• Automated extraction methods are needed due to text complexity and data volume.

Purpose of the Study:

• To develop an automated information extraction model for bacterial associations.
• To leverage transfer learning and fine-tuning for improved prediction accuracy.
• To aid in understanding microbial community structures.

Main Methods:

• A BioBERT-based information extraction model was developed.
• Transfer learning from public datasets was utilized.
• A specialized sentence corpus was created for fine-tuning the model.

Main Results:

• The transfer-learned and fine-tuned model significantly outperformed other variations.
• The model demonstrated superior performance compared to BioGPT-trained models.
• A case study validated the model's utility with experimental data.

Conclusions:

• Transfer learning is applicable to extracting inter-bacterial relationships in life sciences.
• Fine-tuning with limited data can further enhance model performance.
• The model and datasets contribute to medical informatics and bioinformatics.